Modeling finer microstructure of random heterogeneous porous materials by fusing 2D high- and 3D low- resolution images

Abstract

Modeling the finer microstructure is crucial for studying the microscopic and macroscopic characteristics of random heterogeneous porous materials. Multi-scale reconstruction is a feasible way to model finer microstructures. However, most of the traditional methods focus on reconstructing a single type of low-level features, while the neural network-based methods are heavily dependent on the dataset. This paper proposes a novel multi-scale reconstruction framework based on Neural Network-based Transform Mapping (NNTM) to modeling the finer microstructure. The number of network layer and input noise scale are determined by the pixel-size ratio between LRI and HRI. This allows the network to fit the corresponding transformation sub-mapping at different feature scales, thus effectively and accurately fusing the multi-scale information of 2D HRI and 3D LRI. More notably, without the training on a large dataset, and providing only a 2D HRI and a 3D LRI of the same sample, NNTM enables the modeling of high-resolution finer microstructure. Experiments were conducted on rock materials with different pore morphologies, and the morphological characteristics and physical properties were compared to the original high-resolution structure. The qualitative and quantitative evaluation criteria indicate a high consistency between the reconstructed finer microstructure and the original structure.